fs/gfs2/glock.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/sched.h>
  13 #include <linux/slab.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/buffer_head.h>
  16 #include <linux/delay.h>
  17 #include <linux/sort.h>
  18 #include <linux/hash.h>
  19 #include <linux/jhash.h>
  20 #include <linux/kallsyms.h>
  21 #include <linux/gfs2_ondisk.h>
  22 #include <linux/list.h>
  23 #include <linux/wait.h>
  24 #include <linux/module.h>
  25 #include <linux/uaccess.h>
  26 #include <linux/seq_file.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/kthread.h>
  29 #include <linux/freezer.h>
  30 #include <linux/workqueue.h>
  31 #include <linux/jiffies.h>
  32 #include <linux/rcupdate.h>
  33 #include <linux/rculist_bl.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/percpu.h>
  36 #include <linux/list_sort.h>
  37 #include <linux/lockref.h>
  38 #include <linux/rhashtable.h>
  39
  40 #include "gfs2.h"
  41 #include "incore.h"
  42 #include "glock.h"
  43 #include "glops.h"
  44 #include "inode.h"
  45 #include "lops.h"
  46 #include "meta_io.h"
  47 #include "quota.h"
  48 #include "super.h"
  49 #include "util.h"
  50 #include "bmap.h"
  51 #define CREATE_TRACE_POINTS
  52 #include "trace_gfs2.h"
  53
  54 struct gfs2_glock_iter {
  55         struct gfs2_sbd *sdp;           /* incore superblock           */
  56         struct rhashtable_iter hti;     /* rhashtable iterator         */
  57         struct gfs2_glock *gl;          /* current glock struct        */
  58         loff_t last_pos;                /* last position               */
  59 };
  60
  61 typedef void (*glock_examiner) (struct gfs2_glock * gl);
  62
  63 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target);
  64
  65 static struct dentry *gfs2_root;
  66 static struct workqueue_struct *glock_workqueue;
  67 struct workqueue_struct *gfs2_delete_workqueue;
  68 static LIST_HEAD(lru_list);
  69 static atomic_t lru_count = ATOMIC_INIT(0);
  70 static DEFINE_SPINLOCK(lru_lock);
  71
  72 #define GFS2_GL_HASH_SHIFT      15
  73 #define GFS2_GL_HASH_SIZE       BIT(GFS2_GL_HASH_SHIFT)
  74
  75 static const struct rhashtable_params ht_parms = {
  76         .nelem_hint = GFS2_GL_HASH_SIZE * 3 / 4,
  77         .key_len = offsetofend(struct lm_lockname, ln_type),
  78         .key_offset = offsetof(struct gfs2_glock, gl_name),
  79         .head_offset = offsetof(struct gfs2_glock, gl_node),
  80 };
  81
  82 static struct rhashtable gl_hash_table;
  83
  84 #define GLOCK_WAIT_TABLE_BITS 12
  85 #define GLOCK_WAIT_TABLE_SIZE (1 << GLOCK_WAIT_TABLE_BITS)
  86 static wait_queue_head_t glock_wait_table[GLOCK_WAIT_TABLE_SIZE] __cacheline_aligned;
  87
  88 struct wait_glock_queue {
  89         struct lm_lockname *name;
  90         wait_queue_entry_t wait;
  91 };
  92
  93 static int glock_wake_function(wait_queue_entry_t *wait, unsigned int mode,
  94                                int sync, void *key)
  95 {
  96         struct wait_glock_queue *wait_glock =
  97                 container_of(wait, struct wait_glock_queue, wait);
  98         struct lm_lockname *wait_name = wait_glock->name;
  99         struct lm_lockname *wake_name = key;
 100
 101         if (wake_name->ln_sbd != wait_name->ln_sbd ||
 102             wake_name->ln_number != wait_name->ln_number ||
 103             wake_name->ln_type != wait_name->ln_type)
 104                 return 0;
 105         return autoremove_wake_function(wait, mode, sync, key);
 106 }
 107
 108 static wait_queue_head_t *glock_waitqueue(struct lm_lockname *name)
 109 {
 110         u32 hash = jhash2((u32 *)name, sizeof(*name) / 4, 0);
 111
 112         return glock_wait_table + hash_32(hash, GLOCK_WAIT_TABLE_BITS);
 113 }
 114
 115 /**
 116  * wake_up_glock  -  Wake up waiters on a glock
 117  * @gl: the glock
 118  */
 119 static void wake_up_glock(struct gfs2_glock *gl)
 120 {
 121         wait_queue_head_t *wq = glock_waitqueue(&gl->gl_name);
 122
 123         if (waitqueue_active(wq))
 124                 __wake_up(wq, TASK_NORMAL, 1, &gl->gl_name);
 125 }
 126
 127 static void gfs2_glock_dealloc(struct rcu_head *rcu)
 128 {
 129         struct gfs2_glock *gl = container_of(rcu, struct gfs2_glock, gl_rcu);
 130
 131         if (gl->gl_ops->go_flags & GLOF_ASPACE) {
 132                 kmem_cache_free(gfs2_glock_aspace_cachep, gl);
 133         } else {
 134                 kfree(gl->gl_lksb.sb_lvbptr);
 135                 kmem_cache_free(gfs2_glock_cachep, gl);
 136         }
 137 }
 138
 139 void gfs2_glock_free(struct gfs2_glock *gl)
 140 {
 141         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 142
 143         rhashtable_remove_fast(&gl_hash_table, &gl->gl_node, ht_parms);
 144         smp_mb();
 145         wake_up_glock(gl);
 146         call_rcu(&gl->gl_rcu, gfs2_glock_dealloc);
 147         if (atomic_dec_and_test(&sdp->sd_glock_disposal))
 148                 wake_up(&sdp->sd_glock_wait);
 149 }
 150
 151 /**
 152  * gfs2_glock_hold() - increment reference count on glock
 153  * @gl: The glock to hold
 154  *
 155  */
 156
 157 void gfs2_glock_hold(struct gfs2_glock *gl)
 158 {
 159         GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 160         lockref_get(&gl->gl_lockref);
 161 }
 162
 163 /**
 164  * demote_ok - Check to see if it's ok to unlock a glock
 165  * @gl: the glock
 166  *
 167  * Returns: 1 if it's ok
 168  */
 169
 170 static int demote_ok(const struct gfs2_glock *gl)
 171 {
 172         const struct gfs2_glock_operations *glops = gl->gl_ops;
 173
 174         if (gl->gl_state == LM_ST_UNLOCKED)
 175                 return 0;
 176         if (!list_empty(&gl->gl_holders))
 177                 return 0;
 178         if (glops->go_demote_ok)
 179                 return glops->go_demote_ok(gl);
 180         return 1;
 181 }
 182
 183
 184 void gfs2_glock_add_to_lru(struct gfs2_glock *gl)
 185 {
 186         spin_lock(&lru_lock);
 187
 188         if (!list_empty(&gl->gl_lru))
 189                 list_del_init(&gl->gl_lru);
 190         else
 191                 atomic_inc(&lru_count);
 192
 193         list_add_tail(&gl->gl_lru, &lru_list);
 194         set_bit(GLF_LRU, &gl->gl_flags);
 195         spin_unlock(&lru_lock);
 196 }
 197
 198 static void gfs2_glock_remove_from_lru(struct gfs2_glock *gl)
 199 {
 200         if (!(gl->gl_ops->go_flags & GLOF_LRU))
 201                 return;
 202
 203         spin_lock(&lru_lock);
 204         if (!list_empty(&gl->gl_lru)) {
 205                 list_del_init(&gl->gl_lru);
 206                 atomic_dec(&lru_count);
 207                 clear_bit(GLF_LRU, &gl->gl_flags);
 208         }
 209         spin_unlock(&lru_lock);
 210 }
 211
 212 /*
 213  * Enqueue the glock on the work queue.  Passes one glock reference on to the
 214  * work queue.
 215  */
 216 static void __gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 217         if (!queue_delayed_work(glock_workqueue, &gl->gl_work, delay)) {
 218                 /*
 219                  * We are holding the lockref spinlock, and the work was still
 220                  * queued above.  The queued work (glock_work_func) takes that
 221                  * spinlock before dropping its glock reference(s), so it
 222                  * cannot have dropped them in the meantime.
 223                  */
 224                 GLOCK_BUG_ON(gl, gl->gl_lockref.count < 2);
 225                 gl->gl_lockref.count--;
 226         }
 227 }
 228
 229 static void gfs2_glock_queue_work(struct gfs2_glock *gl, unsigned long delay) {
 230         spin_lock(&gl->gl_lockref.lock);
 231         __gfs2_glock_queue_work(gl, delay);
 232         spin_unlock(&gl->gl_lockref.lock);
 233 }
 234
 235 static void __gfs2_glock_put(struct gfs2_glock *gl)
 236 {
 237         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 238         struct address_space *mapping = gfs2_glock2aspace(gl);
 239
 240         lockref_mark_dead(&gl->gl_lockref);
 241
 242         gfs2_glock_remove_from_lru(gl);
 243         spin_unlock(&gl->gl_lockref.lock);
 244         GLOCK_BUG_ON(gl, !list_empty(&gl->gl_holders));
 245         GLOCK_BUG_ON(gl, mapping && mapping->nrpages);
 246         trace_gfs2_glock_put(gl);
 247         sdp->sd_lockstruct.ls_ops->lm_put_lock(gl);
 248 }
 249
 250 /*
 251  * Cause the glock to be put in work queue context.
 252  */
 253 void gfs2_glock_queue_put(struct gfs2_glock *gl)
 254 {
 255         gfs2_glock_queue_work(gl, 0);
 256 }
 257
 258 /**
 259  * gfs2_glock_put() - Decrement reference count on glock
 260  * @gl: The glock to put
 261  *
 262  */
 263
 264 void gfs2_glock_put(struct gfs2_glock *gl)
 265 {
 266         if (lockref_put_or_lock(&gl->gl_lockref))
 267                 return;
 268
 269         __gfs2_glock_put(gl);
 270 }
 271
 272 /**
 273  * may_grant - check if its ok to grant a new lock
 274  * @gl: The glock
 275  * @gh: The lock request which we wish to grant
 276  *
 277  * Returns: true if its ok to grant the lock
 278  */
 279
 280 static inline int may_grant(const struct gfs2_glock *gl, const struct gfs2_holder *gh)
 281 {
 282         const struct gfs2_holder *gh_head = list_entry(gl->gl_holders.next, const struct gfs2_holder, gh_list);
 283         if ((gh->gh_state == LM_ST_EXCLUSIVE ||
 284              gh_head->gh_state == LM_ST_EXCLUSIVE) && gh != gh_head)
 285                 return 0;
 286         if (gl->gl_state == gh->gh_state)
 287                 return 1;
 288         if (gh->gh_flags & GL_EXACT)
 289                 return 0;
 290         if (gl->gl_state == LM_ST_EXCLUSIVE) {
 291                 if (gh->gh_state == LM_ST_SHARED && gh_head->gh_state == LM_ST_SHARED)
 292                         return 1;
 293                 if (gh->gh_state == LM_ST_DEFERRED && gh_head->gh_state == LM_ST_DEFERRED)
 294                         return 1;
 295         }
 296         if (gl->gl_state != LM_ST_UNLOCKED && (gh->gh_flags & LM_FLAG_ANY))
 297                 return 1;
 298         return 0;
 299 }
 300
 301 static void gfs2_holder_wake(struct gfs2_holder *gh)
 302 {
 303         clear_bit(HIF_WAIT, &gh->gh_iflags);
 304         smp_mb__after_atomic();
 305         wake_up_bit(&gh->gh_iflags, HIF_WAIT);
 306 }
 307
 308 /**
 309  * do_error - Something unexpected has happened during a lock request
 310  *
 311  */
 312
 313 static void do_error(struct gfs2_glock *gl, const int ret)
 314 {
 315         struct gfs2_holder *gh, *tmp;
 316
 317         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 318                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 319                         continue;
 320                 if (ret & LM_OUT_ERROR)
 321                         gh->gh_error = -EIO;
 322                 else if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))
 323                         gh->gh_error = GLR_TRYFAILED;
 324                 else
 325                         continue;
 326                 list_del_init(&gh->gh_list);
 327                 trace_gfs2_glock_queue(gh, 0);
 328                 gfs2_holder_wake(gh);
 329         }
 330 }
 331
 332 /**
 333  * do_promote - promote as many requests as possible on the current queue
 334  * @gl: The glock
 335  *
 336  * Returns: 1 if there is a blocked holder at the head of the list, or 2
 337  *          if a type specific operation is underway.
 338  */
 339
 340 static int do_promote(struct gfs2_glock *gl)
 341 __releases(&gl->gl_lockref.lock)
 342 __acquires(&gl->gl_lockref.lock)
 343 {
 344         const struct gfs2_glock_operations *glops = gl->gl_ops;
 345         struct gfs2_holder *gh, *tmp;
 346         int ret;
 347
 348 restart:
 349         list_for_each_entry_safe(gh, tmp, &gl->gl_holders, gh_list) {
 350                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 351                         continue;
 352                 if (may_grant(gl, gh)) {
 353                         if (gh->gh_list.prev == &gl->gl_holders &&
 354                             glops->go_lock) {
 355                                 spin_unlock(&gl->gl_lockref.lock);
 356                                 /* FIXME: eliminate this eventually */
 357                                 ret = glops->go_lock(gh);
 358                                 spin_lock(&gl->gl_lockref.lock);
 359                                 if (ret) {
 360                                         if (ret == 1)
 361                                                 return 2;
 362                                         gh->gh_error = ret;
 363                                         list_del_init(&gh->gh_list);
 364                                         trace_gfs2_glock_queue(gh, 0);
 365                                         gfs2_holder_wake(gh);
 366                                         goto restart;
 367                                 }
 368                                 set_bit(HIF_HOLDER, &gh->gh_iflags);
 369                                 trace_gfs2_promote(gh, 1);
 370                                 gfs2_holder_wake(gh);
 371                                 goto restart;
 372                         }
 373                         set_bit(HIF_HOLDER, &gh->gh_iflags);
 374                         trace_gfs2_promote(gh, 0);
 375                         gfs2_holder_wake(gh);
 376                         continue;
 377                 }
 378                 if (gh->gh_list.prev == &gl->gl_holders)
 379                         return 1;
 380                 do_error(gl, 0);
 381                 break;
 382         }
 383         return 0;
 384 }
 385
 386 /**
 387  * find_first_waiter - find the first gh that's waiting for the glock
 388  * @gl: the glock
 389  */
 390
 391 static inline struct gfs2_holder *find_first_waiter(const struct gfs2_glock *gl)
 392 {
 393         struct gfs2_holder *gh;
 394
 395         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
 396                 if (!test_bit(HIF_HOLDER, &gh->gh_iflags))
 397                         return gh;
 398         }
 399         return NULL;
 400 }
 401
 402 /**
 403  * state_change - record that the glock is now in a different state
 404  * @gl: the glock
 405  * @new_state the new state
 406  *
 407  */
 408
 409 static void state_change(struct gfs2_glock *gl, unsigned int new_state)
 410 {
 411         int held1, held2;
 412
 413         held1 = (gl->gl_state != LM_ST_UNLOCKED);
 414         held2 = (new_state != LM_ST_UNLOCKED);
 415
 416         if (held1 != held2) {
 417                 GLOCK_BUG_ON(gl, __lockref_is_dead(&gl->gl_lockref));
 418                 if (held2)
 419                         gl->gl_lockref.count++;
 420                 else
 421                         gl->gl_lockref.count--;
 422         }
 423         if (held1 && held2 && list_empty(&gl->gl_holders))
 424                 clear_bit(GLF_QUEUED, &gl->gl_flags);
 425
 426         if (new_state != gl->gl_target)
 427                 /* shorten our minimum hold time */
 428                 gl->gl_hold_time = max(gl->gl_hold_time - GL_GLOCK_HOLD_DECR,
 429                                        GL_GLOCK_MIN_HOLD);
 430         gl->gl_state = new_state;
 431         gl->gl_tchange = jiffies;
 432 }
 433
 434 static void gfs2_demote_wake(struct gfs2_glock *gl)
 435 {
 436         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 437         clear_bit(GLF_DEMOTE, &gl->gl_flags);
 438         smp_mb__after_atomic();
 439         wake_up_bit(&gl->gl_flags, GLF_DEMOTE);
 440 }
 441
 442 /**
 443  * finish_xmote - The DLM has replied to one of our lock requests
 444  * @gl: The glock
 445  * @ret: The status from the DLM
 446  *
 447  */
 448
 449 static void finish_xmote(struct gfs2_glock *gl, unsigned int ret)
 450 {
 451         const struct gfs2_glock_operations *glops = gl->gl_ops;
 452         struct gfs2_holder *gh;
 453         unsigned state = ret & LM_OUT_ST_MASK;
 454         int rv;
 455
 456         spin_lock(&gl->gl_lockref.lock);
 457         trace_gfs2_glock_state_change(gl, state);
 458         state_change(gl, state);
 459         gh = find_first_waiter(gl);
 460
 461         /* Demote to UN request arrived during demote to SH or DF */
 462         if (test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags) &&
 463             state != LM_ST_UNLOCKED && gl->gl_demote_state == LM_ST_UNLOCKED)
 464                 gl->gl_target = LM_ST_UNLOCKED;
 465
 466         /* Check for state != intended state */
 467         if (unlikely(state != gl->gl_target)) {
 468                 if (gh && !test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags)) {
 469                         /* move to back of queue and try next entry */
 470                         if (ret & LM_OUT_CANCELED) {
 471                                 if ((gh->gh_flags & LM_FLAG_PRIORITY) == 0)
 472                                         list_move_tail(&gh->gh_list, &gl->gl_holders);
 473                                 gh = find_first_waiter(gl);
 474                                 gl->gl_target = gh->gh_state;
 475                                 goto retry;
 476                         }
 477                         /* Some error or failed "try lock" - report it */
 478                         if ((ret & LM_OUT_ERROR) ||
 479                             (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
 480                                 gl->gl_target = gl->gl_state;
 481                                 do_error(gl, ret);
 482                                 goto out;
 483                         }
 484                 }
 485                 switch(state) {
 486                 /* Unlocked due to conversion deadlock, try again */
 487                 case LM_ST_UNLOCKED:
 488 retry:
 489                         do_xmote(gl, gh, gl->gl_target);
 490                         break;
 491                 /* Conversion fails, unlock and try again */
 492                 case LM_ST_SHARED:
 493                 case LM_ST_DEFERRED:
 494                         do_xmote(gl, gh, LM_ST_UNLOCKED);
 495                         break;
 496                 default: /* Everything else */
 497                         fs_err(gl->gl_name.ln_sbd, "wanted %u got %u\n",
 498                                gl->gl_target, state);
 499                         GLOCK_BUG_ON(gl, 1);
 500                 }
 501                 spin_unlock(&gl->gl_lockref.lock);
 502                 return;
 503         }
 504
 505         /* Fast path - we got what we asked for */
 506         if (test_and_clear_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags))
 507                 gfs2_demote_wake(gl);
 508         if (state != LM_ST_UNLOCKED) {
 509                 if (glops->go_xmote_bh) {
 510                         spin_unlock(&gl->gl_lockref.lock);
 511                         rv = glops->go_xmote_bh(gl, gh);
 512                         spin_lock(&gl->gl_lockref.lock);
 513                         if (rv) {
 514                                 do_error(gl, rv);
 515                                 goto out;
 516                         }
 517                 }
 518                 rv = do_promote(gl);
 519                 if (rv == 2)
 520                         goto out_locked;
 521         }
 522 out:
 523         clear_bit(GLF_LOCK, &gl->gl_flags);
 524 out_locked:
 525         spin_unlock(&gl->gl_lockref.lock);
 526 }
 527
 528 /**
 529  * do_xmote - Calls the DLM to change the state of a lock
 530  * @gl: The lock state
 531  * @gh: The holder (only for promotes)
 532  * @target: The target lock state
 533  *
 534  */
 535
 536 static void do_xmote(struct gfs2_glock *gl, struct gfs2_holder *gh, unsigned int target)
 537 __releases(&gl->gl_lockref.lock)
 538 __acquires(&gl->gl_lockref.lock)
 539 {
 540         const struct gfs2_glock_operations *glops = gl->gl_ops;
 541         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 542         unsigned int lck_flags = (unsigned int)(gh ? gh->gh_flags : 0);
 543         int ret;
 544
 545         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)) &&
 546             target != LM_ST_UNLOCKED)
 547                 return;
 548         lck_flags &= (LM_FLAG_TRY | LM_FLAG_TRY_1CB | LM_FLAG_NOEXP |
 549                       LM_FLAG_PRIORITY);
 550         GLOCK_BUG_ON(gl, gl->gl_state == target);
 551         GLOCK_BUG_ON(gl, gl->gl_state == gl->gl_target);
 552         if ((target == LM_ST_UNLOCKED || target == LM_ST_DEFERRED) &&
 553             glops->go_inval) {
 554                 set_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 555                 do_error(gl, 0); /* Fail queued try locks */
 556         }
 557         gl->gl_req = target;
 558         set_bit(GLF_BLOCKING, &gl->gl_flags);
 559         if ((gl->gl_req == LM_ST_UNLOCKED) ||
 560             (gl->gl_state == LM_ST_EXCLUSIVE) ||
 561             (lck_flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB)))
 562                 clear_bit(GLF_BLOCKING, &gl->gl_flags);
 563         spin_unlock(&gl->gl_lockref.lock);
 564         if (glops->go_sync)
 565                 glops->go_sync(gl);
 566         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
 567                 glops->go_inval(gl, target == LM_ST_DEFERRED ? 0 : DIO_METADATA);
 568         clear_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags);
 569
 570         gfs2_glock_hold(gl);
 571         if (sdp->sd_lockstruct.ls_ops->lm_lock) {
 572                 /* lock_dlm */
 573                 ret = sdp->sd_lockstruct.ls_ops->lm_lock(gl, target, lck_flags);
 574                 if (ret == -EINVAL && gl->gl_target == LM_ST_UNLOCKED &&
 575                     target == LM_ST_UNLOCKED &&
 576                     test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags)) {
 577                         finish_xmote(gl, target);
 578                         gfs2_glock_queue_work(gl, 0);
 579                 }
 580                 else if (ret) {
 581                         fs_err(sdp, "lm_lock ret %d\n", ret);
 582                         GLOCK_BUG_ON(gl, !test_bit(SDF_SHUTDOWN,
 583                                                    &sdp->sd_flags));
 584                 }
 585         } else { /* lock_nolock */
 586                 finish_xmote(gl, target);
 587                 gfs2_glock_queue_work(gl, 0);
 588         }
 589
 590         spin_lock(&gl->gl_lockref.lock);
 591 }
 592
 593 /**
 594  * find_first_holder - find the first "holder" gh
 595  * @gl: the glock
 596  */
 597
 598 static inline struct gfs2_holder *find_first_holder(const struct gfs2_glock *gl)
 599 {
 600         struct gfs2_holder *gh;
 601
 602         if (!list_empty(&gl->gl_holders)) {
 603                 gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
 604                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
 605                         return gh;
 606         }
 607         return NULL;
 608 }
 609
 610 /**
 611  * run_queue - do all outstanding tasks related to a glock
 612  * @gl: The glock in question
 613  * @nonblock: True if we must not block in run_queue
 614  *
 615  */
 616
 617 static void run_queue(struct gfs2_glock *gl, const int nonblock)
 618 __releases(&gl->gl_lockref.lock)
 619 __acquires(&gl->gl_lockref.lock)
 620 {
 621         struct gfs2_holder *gh = NULL;
 622         int ret;
 623
 624         if (test_and_set_bit(GLF_LOCK, &gl->gl_flags))
 625                 return;
 626
 627         GLOCK_BUG_ON(gl, test_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags));
 628
 629         if (test_bit(GLF_DEMOTE, &gl->gl_flags) &&
 630             gl->gl_demote_state != gl->gl_state) {
 631                 if (find_first_holder(gl))
 632                         goto out_unlock;
 633                 if (nonblock)
 634                         goto out_sched;
 635                 set_bit(GLF_DEMOTE_IN_PROGRESS, &gl->gl_flags);
 636                 GLOCK_BUG_ON(gl, gl->gl_demote_state == LM_ST_EXCLUSIVE);
 637                 gl->gl_target = gl->gl_demote_state;
 638         } else {
 639                 if (test_bit(GLF_DEMOTE, &gl->gl_flags))
 640                         gfs2_demote_wake(gl);
 641                 ret = do_promote(gl);
 642                 if (ret == 0)
 643                         goto out_unlock;
 644                 if (ret == 2)
 645                         goto out;
 646                 gh = find_first_waiter(gl);
 647                 gl->gl_target = gh->gh_state;
 648                 if (!(gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)))
 649                         do_error(gl, 0); /* Fail queued try locks */
 650         }
 651         do_xmote(gl, gh, gl->gl_target);
 652 out:
 653         return;
 654
 655 out_sched:
 656         clear_bit(GLF_LOCK, &gl->gl_flags);
 657         smp_mb__after_atomic();
 658         gl->gl_lockref.count++;
 659         __gfs2_glock_queue_work(gl, 0);
 660         return;
 661
 662 out_unlock:
 663         clear_bit(GLF_LOCK, &gl->gl_flags);
 664         smp_mb__after_atomic();
 665         return;
 666 }
 667
 668 static void delete_work_func(struct work_struct *work)
 669 {
 670         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_delete);
 671         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
 672         struct inode *inode;
 673         u64 no_addr = gl->gl_name.ln_number;
 674
 675         /* If someone's using this glock to create a new dinode, the block must
 676            have been freed by another node, then re-used, in which case our
 677            iopen callback is too late after the fact. Ignore it. */
 678         if (test_bit(GLF_INODE_CREATING, &gl->gl_flags))
 679                 goto out;
 680
 681         inode = gfs2_lookup_by_inum(sdp, no_addr, NULL, GFS2_BLKST_UNLINKED);
 682         if (inode && !IS_ERR(inode)) {
 683                 d_prune_aliases(inode);
 684                 iput(inode);
 685         }
 686 out:
 687         gfs2_glock_put(gl);
 688 }
 689
 690 static void glock_work_func(struct work_struct *work)
 691 {
 692         unsigned long delay = 0;
 693         struct gfs2_glock *gl = container_of(work, struct gfs2_glock, gl_work.work);
 694         unsigned int drop_refs = 1;
 695
 696         if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
 697                 finish_xmote(gl, gl->gl_reply);
 698                 drop_refs++;
 699         }
 700         spin_lock(&gl->gl_lockref.lock);
 701         if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
 702             gl->gl_state != LM_ST_UNLOCKED &&
 703             gl->gl_demote_state != LM_ST_EXCLUSIVE) {
 704                 unsigned long holdtime, now = jiffies;
 705
 706                 holdtime = gl->gl_tchange + gl->gl_hold_time;
 707                 if (time_before(now, holdtime))
 708                         delay = holdtime - now;
 709
 710                 if (!delay) {
 711                         clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags);
 712                         set_bit(GLF_DEMOTE, &gl->gl_flags);
 713                 }
 714         }
 715         run_queue(gl, 0);
 716         if (delay) {
 717                 /* Keep one glock reference for the work we requeue. */
 718                 drop_refs--;
 719                 if (gl->gl_name.ln_type != LM_TYPE_INODE)
 720                         delay = 0;
 721                 __gfs2_glock_queue_work(gl, delay);
 722         }
 723
 724         /*
 725          * Drop the remaining glock references manually here. (Mind that
 726          * __gfs2_glock_queue_work depends on the lockref spinlock begin held
 727          * here as well.)
 728          */
 729         gl->gl_lockref.count -= drop_refs;
 730         if (!gl->gl_lockref.count) {
 731                 __gfs2_glock_put(gl);
 732                 return;
 733         }
 734         spin_unlock(&gl->gl_lockref.lock);
 735 }
 736
 737 static struct gfs2_glock *find_insert_glock(struct lm_lockname *name,
 738                                             struct gfs2_glock *new)
 739 {
 740         struct wait_glock_queue wait;
 741         wait_queue_head_t *wq = glock_waitqueue(name);
 742         struct gfs2_glock *gl;
 743
 744         wait.name = name;
 745         init_wait(&wait.wait);
 746         wait.wait.func = glock_wake_function;
 747
 748 again:
 749         prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
 750         rcu_read_lock();
 751         if (new) {
 752                 gl = rhashtable_lookup_get_insert_fast(&gl_hash_table,
 753                         &new->gl_node, ht_parms);
 754                 if (IS_ERR(gl))
 755                         goto out;
 756         } else {
 757                 gl = rhashtable_lookup_fast(&gl_hash_table,
 758                         name, ht_parms);
 759         }
 760         if (gl && !lockref_get_not_dead(&gl->gl_lockref)) {
 761                 rcu_read_unlock();
 762                 schedule();
 763                 goto again;
 764         }
 765 out:
 766         rcu_read_unlock();
 767         finish_wait(wq, &wait.wait);
 768         return gl;
 769 }
 770
 771 /**
 772  * gfs2_glock_get() - Get a glock, or create one if one doesn't exist
 773  * @sdp: The GFS2 superblock
 774  * @number: the lock number
 775  * @glops: The glock_operations to use
 776  * @create: If 0, don't create the glock if it doesn't exist
 777  * @glp: the glock is returned here
 778  *
 779  * This does not lock a glock, just finds/creates structures for one.
 780  *
 781  * Returns: errno
 782  */
 783
 784 int gfs2_glock_get(struct gfs2_sbd *sdp, u64 number,
 785                    const struct gfs2_glock_operations *glops, int create,
 786                    struct gfs2_glock **glp)
 787 {
 788         struct super_block *s = sdp->sd_vfs;
 789         struct lm_lockname name = { .ln_number = number,
 790                                     .ln_type = glops->go_type,
 791                                     .ln_sbd = sdp };
 792         struct gfs2_glock *gl, *tmp;
 793         struct address_space *mapping;
 794         struct kmem_cache *cachep;
 795         int ret = 0;
 796
 797         gl = find_insert_glock(&name, NULL);
 798         if (gl) {
 799                 *glp = gl;
 800                 return 0;
 801         }
 802         if (!create)
 803                 return -ENOENT;
 804
 805         if (glops->go_flags & GLOF_ASPACE)
 806                 cachep = gfs2_glock_aspace_cachep;
 807         else
 808                 cachep = gfs2_glock_cachep;
 809         gl = kmem_cache_alloc(cachep, GFP_NOFS);
 810         if (!gl)
 811                 return -ENOMEM;
 812
 813         memset(&gl->gl_lksb, 0, sizeof(struct dlm_lksb));
 814
 815         if (glops->go_flags & GLOF_LVB) {
 816                 gl->gl_lksb.sb_lvbptr = kzalloc(GFS2_MIN_LVB_SIZE, GFP_NOFS);
 817                 if (!gl->gl_lksb.sb_lvbptr) {
 818                         kmem_cache_free(cachep, gl);
 819                         return -ENOMEM;
 820                 }
 821         }
 822
 823         atomic_inc(&sdp->sd_glock_disposal);
 824         gl->gl_node.next = NULL;
 825         gl->gl_flags = 0;
 826         gl->gl_name = name;
 827         gl->gl_lockref.count = 1;
 828         gl->gl_state = LM_ST_UNLOCKED;
 829         gl->gl_target = LM_ST_UNLOCKED;
 830         gl->gl_demote_state = LM_ST_EXCLUSIVE;
 831         gl->gl_ops = glops;
 832         gl->gl_dstamp = 0;
 833         preempt_disable();
 834         /* We use the global stats to estimate the initial per-glock stats */
 835         gl->gl_stats = this_cpu_ptr(sdp->sd_lkstats)->lkstats[glops->go_type];
 836         preempt_enable();
 837         gl->gl_stats.stats[GFS2_LKS_DCOUNT] = 0;
 838         gl->gl_stats.stats[GFS2_LKS_QCOUNT] = 0;
 839         gl->gl_tchange = jiffies;
 840         gl->gl_object = NULL;
 841         gl->gl_hold_time = GL_GLOCK_DFT_HOLD;
 842         INIT_DELAYED_WORK(&gl->gl_work, glock_work_func);
 843         INIT_WORK(&gl->gl_delete, delete_work_func);
 844
 845         mapping = gfs2_glock2aspace(gl);
 846         if (mapping) {
 847                 mapping->a_ops = &gfs2_meta_aops;
 848                 mapping->host = s->s_bdev->bd_inode;
 849                 mapping->flags = 0;
 850                 mapping_set_gfp_mask(mapping, GFP_NOFS);
 851                 mapping->private_data = NULL;
 852                 mapping->writeback_index = 0;
 853         }
 854
 855         tmp = find_insert_glock(&name, gl);
 856         if (!tmp) {
 857                 *glp = gl;
 858                 goto out;
 859         }
 860         if (IS_ERR(tmp)) {
 861                 ret = PTR_ERR(tmp);
 862                 goto out_free;
 863         }
 864         *glp = tmp;
 865
 866 out_free:
 867         kfree(gl->gl_lksb.sb_lvbptr);
 868         kmem_cache_free(cachep, gl);
 869         atomic_dec(&sdp->sd_glock_disposal);
 870
 871 out:
 872         return ret;
 873 }
 874
 875 /**
 876  * gfs2_holder_init - initialize a struct gfs2_holder in the default way
 877  * @gl: the glock
 878  * @state: the state we're requesting
 879  * @flags: the modifier flags
 880  * @gh: the holder structure
 881  *
 882  */
 883
 884 void gfs2_holder_init(struct gfs2_glock *gl, unsigned int state, u16 flags,
 885                       struct gfs2_holder *gh)
 886 {
 887         INIT_LIST_HEAD(&gh->gh_list);
 888         gh->gh_gl = gl;
 889         gh->gh_ip = _RET_IP_;
 890         gh->gh_owner_pid = get_pid(task_pid(current));
 891         gh->gh_state = state;
 892         gh->gh_flags = flags;
 893         gh->gh_error = 0;
 894         gh->gh_iflags = 0;
 895         gfs2_glock_hold(gl);
 896 }
 897
 898 /**
 899  * gfs2_holder_reinit - reinitialize a struct gfs2_holder so we can requeue it
 900  * @state: the state we're requesting
 901  * @flags: the modifier flags
 902  * @gh: the holder structure
 903  *
 904  * Don't mess with the glock.
 905  *
 906  */
 907
 908 void gfs2_holder_reinit(unsigned int state, u16 flags, struct gfs2_holder *gh)
 909 {
 910         gh->gh_state = state;
 911         gh->gh_flags = flags;
 912         gh->gh_iflags = 0;
 913         gh->gh_ip = _RET_IP_;
 914         put_pid(gh->gh_owner_pid);
 915         gh->gh_owner_pid = get_pid(task_pid(current));
 916 }
 917
 918 /**
 919  * gfs2_holder_uninit - uninitialize a holder structure (drop glock reference)
 920  * @gh: the holder structure
 921  *
 922  */
 923
 924 void gfs2_holder_uninit(struct gfs2_holder *gh)
 925 {
 926         put_pid(gh->gh_owner_pid);
 927         gfs2_glock_put(gh->gh_gl);
 928         gfs2_holder_mark_uninitialized(gh);
 929         gh->gh_ip = 0;
 930 }
 931
 932 /**
 933  * gfs2_glock_wait - wait on a glock acquisition
 934  * @gh: the glock holder
 935  *
 936  * Returns: 0 on success
 937  */
 938
 939 int gfs2_glock_wait(struct gfs2_holder *gh)
 940 {
 941         unsigned long time1 = jiffies;
 942
 943         might_sleep();
 944         wait_on_bit(&gh->gh_iflags, HIF_WAIT, TASK_UNINTERRUPTIBLE);
 945         if (time_after(jiffies, time1 + HZ)) /* have we waited > a second? */
 946                 /* Lengthen the minimum hold time. */
 947                 gh->gh_gl->gl_hold_time = min(gh->gh_gl->gl_hold_time +
 948                                               GL_GLOCK_HOLD_INCR,
 949                                               GL_GLOCK_MAX_HOLD);
 950         return gh->gh_error;
 951 }
 952
 953 /**
 954  * handle_callback - process a demote request
 955  * @gl: the glock
 956  * @state: the state the caller wants us to change to
 957  *
 958  * There are only two requests that we are going to see in actual
 959  * practise: LM_ST_SHARED and LM_ST_UNLOCKED
 960  */
 961
 962 static void handle_callback(struct gfs2_glock *gl, unsigned int state,
 963                             unsigned long delay, bool remote)
 964 {
 965         int bit = delay ? GLF_PENDING_DEMOTE : GLF_DEMOTE;
 966
 967         set_bit(bit, &gl->gl_flags);
 968         if (gl->gl_demote_state == LM_ST_EXCLUSIVE) {
 969                 gl->gl_demote_state = state;
 970                 gl->gl_demote_time = jiffies;
 971         } else if (gl->gl_demote_state != LM_ST_UNLOCKED &&
 972                         gl->gl_demote_state != state) {
 973                 gl->gl_demote_state = LM_ST_UNLOCKED;
 974         }
 975         if (gl->gl_ops->go_callback)
 976                 gl->gl_ops->go_callback(gl, remote);
 977         trace_gfs2_demote_rq(gl, remote);
 978 }
 979
 980 void gfs2_print_dbg(struct seq_file *seq, const char *fmt, ...)
 981 {
 982         struct va_format vaf;
 983         va_list args;
 984
 985         va_start(args, fmt);
 986
 987         if (seq) {
 988                 seq_vprintf(seq, fmt, args);
 989         } else {
 990                 vaf.fmt = fmt;
 991                 vaf.va = &args;
 992
 993                 pr_err("%pV", &vaf);
 994         }
 995
 996         va_end(args);
 997 }
 998
 999 /**
1000  * add_to_queue - Add a holder to the wait queue (but look for recursion)
1001  * @gh: the holder structure to add
1002  *
1003  * Eventually we should move the recursive locking trap to a
1004  * debugging option or something like that. This is the fast
1005  * path and needs to have the minimum number of distractions.
1006  *
1007  */
1008
1009 static inline void add_to_queue(struct gfs2_holder *gh)
1010 __releases(&gl->gl_lockref.lock)
1011 __acquires(&gl->gl_lockref.lock)
1012 {
1013         struct gfs2_glock *gl = gh->gh_gl;
1014         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1015         struct list_head *insert_pt = NULL;
1016         struct gfs2_holder *gh2;
1017         int try_futile = 0;
1018
1019         BUG_ON(gh->gh_owner_pid == NULL);
1020         if (test_and_set_bit(HIF_WAIT, &gh->gh_iflags))
1021                 BUG();
1022
1023         if (gh->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB)) {
1024                 if (test_bit(GLF_LOCK, &gl->gl_flags))
1025                         try_futile = !may_grant(gl, gh);
1026                 if (test_bit(GLF_INVALIDATE_IN_PROGRESS, &gl->gl_flags))
1027                         goto fail;
1028         }
1029
1030         list_for_each_entry(gh2, &gl->gl_holders, gh_list) {
1031                 if (unlikely(gh2->gh_owner_pid == gh->gh_owner_pid &&
1032                     (gh->gh_gl->gl_ops->go_type != LM_TYPE_FLOCK)))
1033                         goto trap_recursive;
1034                 if (try_futile &&
1035                     !(gh2->gh_flags & (LM_FLAG_TRY | LM_FLAG_TRY_1CB))) {
1036 fail:
1037                         gh->gh_error = GLR_TRYFAILED;
1038                         gfs2_holder_wake(gh);
1039                         return;
1040                 }
1041                 if (test_bit(HIF_HOLDER, &gh2->gh_iflags))
1042                         continue;
1043                 if (unlikely((gh->gh_flags & LM_FLAG_PRIORITY) && !insert_pt))
1044                         insert_pt = &gh2->gh_list;
1045         }
1046         set_bit(GLF_QUEUED, &gl->gl_flags);
1047         trace_gfs2_glock_queue(gh, 1);
1048         gfs2_glstats_inc(gl, GFS2_LKS_QCOUNT);
1049         gfs2_sbstats_inc(gl, GFS2_LKS_QCOUNT);
1050         if (likely(insert_pt == NULL)) {
1051                 list_add_tail(&gh->gh_list, &gl->gl_holders);
1052                 if (unlikely(gh->gh_flags & LM_FLAG_PRIORITY))
1053                         goto do_cancel;
1054                 return;
1055         }
1056         list_add_tail(&gh->gh_list, insert_pt);
1057 do_cancel:
1058         gh = list_entry(gl->gl_holders.next, struct gfs2_holder, gh_list);
1059         if (!(gh->gh_flags & LM_FLAG_PRIORITY)) {
1060                 spin_unlock(&gl->gl_lockref.lock);
1061                 if (sdp->sd_lockstruct.ls_ops->lm_cancel)
1062                         sdp->sd_lockstruct.ls_ops->lm_cancel(gl);
1063                 spin_lock(&gl->gl_lockref.lock);
1064         }
1065         return;
1066
1067 trap_recursive:
1068         fs_err(sdp, "original: %pSR\n", (void *)gh2->gh_ip);
1069         fs_err(sdp, "pid: %d\n", pid_nr(gh2->gh_owner_pid));
1070         fs_err(sdp, "lock type: %d req lock state : %d\n",
1071                gh2->gh_gl->gl_name.ln_type, gh2->gh_state);
1072         fs_err(sdp, "new: %pSR\n", (void *)gh->gh_ip);
1073         fs_err(sdp, "pid: %d\n", pid_nr(gh->gh_owner_pid));
1074         fs_err(sdp, "lock type: %d req lock state : %d\n",
1075                gh->gh_gl->gl_name.ln_type, gh->gh_state);
1076         gfs2_dump_glock(NULL, gl);
1077         BUG();
1078 }
1079
1080 /**
1081  * gfs2_glock_nq - enqueue a struct gfs2_holder onto a glock (acquire a glock)
1082  * @gh: the holder structure
1083  *
1084  * if (gh->gh_flags & GL_ASYNC), this never returns an error
1085  *
1086  * Returns: 0, GLR_TRYFAILED, or errno on failure
1087  */
1088
1089 int gfs2_glock_nq(struct gfs2_holder *gh)
1090 {
1091         struct gfs2_glock *gl = gh->gh_gl;
1092         struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
1093         int error = 0;
1094
1095         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
1096                 return -EIO;
1097
1098         if (test_bit(GLF_LRU, &gl->gl_flags))
1099                 gfs2_glock_remove_from_lru(gl);
1100
1101         spin_lock(&gl->gl_lockref.lock);
1102         add_to_queue(gh);
1103         if (unlikely((LM_FLAG_NOEXP & gh->gh_flags) &&
1104                      test_and_clear_bit(GLF_FROZEN, &gl->gl_flags))) {
1105                 set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1106                 gl->gl_lockref.count++;
1107                 __gfs2_glock_queue_work(gl, 0);
1108         }
1109         run_queue(gl, 1);
1110         spin_unlock(&gl->gl_lockref.lock);
1111
1112         if (!(gh->gh_flags & GL_ASYNC))
1113                 error = gfs2_glock_wait(gh);
1114
1115         return error;
1116 }
1117
1118 /**
1119  * gfs2_glock_poll - poll to see if an async request has been completed
1120  * @gh: the holder
1121  *
1122  * Returns: 1 if the request is ready to be gfs2_glock_wait()ed on
1123  */
1124
1125 int gfs2_glock_poll(struct gfs2_holder *gh)
1126 {
1127         return test_bit(HIF_WAIT, &gh->gh_iflags) ? 0 : 1;
1128 }
1129
1130 /**
1131  * gfs2_glock_dq - dequeue a struct gfs2_holder from a glock (release a glock)
1132  * @gh: the glock holder
1133  *
1134  */
1135
1136 void gfs2_glock_dq(struct gfs2_holder *gh)
1137 {
1138         struct gfs2_glock *gl = gh->gh_gl;
1139         const struct gfs2_glock_operations *glops = gl->gl_ops;
1140         unsigned delay = 0;
1141         int fast_path = 0;
1142
1143         spin_lock(&gl->gl_lockref.lock);
1144         if (gh->gh_flags & GL_NOCACHE)
1145                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1146
1147         list_del_init(&gh->gh_list);
1148         clear_bit(HIF_HOLDER, &gh->gh_iflags);
1149         if (find_first_holder(gl) == NULL) {
1150                 if (glops->go_unlock) {
1151                         GLOCK_BUG_ON(gl, test_and_set_bit(GLF_LOCK, &gl->gl_flags));
1152                         spin_unlock(&gl->gl_lockref.lock);
1153                         glops->go_unlock(gh);
1154                         spin_lock(&gl->gl_lockref.lock);
1155                         clear_bit(GLF_LOCK, &gl->gl_flags);
1156                 }
1157                 if (list_empty(&gl->gl_holders) &&
1158                     !test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1159                     !test_bit(GLF_DEMOTE, &gl->gl_flags))
1160                         fast_path = 1;
1161         }
1162         if (!test_bit(GLF_LFLUSH, &gl->gl_flags) && demote_ok(gl) &&
1163             (glops->go_flags & GLOF_LRU))
1164                 gfs2_glock_add_to_lru(gl);
1165
1166         trace_gfs2_glock_queue(gh, 0);
1167         if (unlikely(!fast_path)) {
1168                 gl->gl_lockref.count++;
1169                 if (test_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
1170                     !test_bit(GLF_DEMOTE, &gl->gl_flags) &&
1171                     gl->gl_name.ln_type == LM_TYPE_INODE)
1172                         delay = gl->gl_hold_time;
1173                 __gfs2_glock_queue_work(gl, delay);
1174         }
1175         spin_unlock(&gl->gl_lockref.lock);
1176 }
1177
1178 void gfs2_glock_dq_wait(struct gfs2_holder *gh)
1179 {
1180         struct gfs2_glock *gl = gh->gh_gl;
1181         gfs2_glock_dq(gh);
1182         might_sleep();
1183         wait_on_bit(&gl->gl_flags, GLF_DEMOTE, TASK_UNINTERRUPTIBLE);
1184 }
1185
1186 /**
1187  * gfs2_glock_dq_uninit - dequeue a holder from a glock and initialize it
1188  * @gh: the holder structure
1189  *
1190  */
1191
1192 void gfs2_glock_dq_uninit(struct gfs2_holder *gh)
1193 {
1194         gfs2_glock_dq(gh);
1195         gfs2_holder_uninit(gh);
1196 }
1197
1198 /**
1199  * gfs2_glock_nq_num - acquire a glock based on lock number
1200  * @sdp: the filesystem
1201  * @number: the lock number
1202  * @glops: the glock operations for the type of glock
1203  * @state: the state to acquire the glock in
1204  * @flags: modifier flags for the acquisition
1205  * @gh: the struct gfs2_holder
1206  *
1207  * Returns: errno
1208  */
1209
1210 int gfs2_glock_nq_num(struct gfs2_sbd *sdp, u64 number,
1211                       const struct gfs2_glock_operations *glops,
1212                       unsigned int state, u16 flags, struct gfs2_holder *gh)
1213 {
1214         struct gfs2_glock *gl;
1215         int error;
1216
1217         error = gfs2_glock_get(sdp, number, glops, CREATE, &gl);
1218         if (!error) {
1219                 error = gfs2_glock_nq_init(gl, state, flags, gh);
1220                 gfs2_glock_put(gl);
1221         }
1222
1223         return error;
1224 }
1225
1226 /**
1227  * glock_compare - Compare two struct gfs2_glock structures for sorting
1228  * @arg_a: the first structure
1229  * @arg_b: the second structure
1230  *
1231  */
1232
1233 static int glock_compare(const void *arg_a, const void *arg_b)
1234 {
1235         const struct gfs2_holder *gh_a = *(const struct gfs2_holder **)arg_a;
1236         const struct gfs2_holder *gh_b = *(const struct gfs2_holder **)arg_b;
1237         const struct lm_lockname *a = &gh_a->gh_gl->gl_name;
1238         const struct lm_lockname *b = &gh_b->gh_gl->gl_name;
1239
1240         if (a->ln_number > b->ln_number)
1241                 return 1;
1242         if (a->ln_number < b->ln_number)
1243                 return -1;
1244         BUG_ON(gh_a->gh_gl->gl_ops->go_type == gh_b->gh_gl->gl_ops->go_type);
1245         return 0;
1246 }
1247
1248 /**
1249  * nq_m_sync - synchonously acquire more than one glock in deadlock free order
1250  * @num_gh: the number of structures
1251  * @ghs: an array of struct gfs2_holder structures
1252  *
1253  * Returns: 0 on success (all glocks acquired),
1254  *          errno on failure (no glocks acquired)
1255  */
1256
1257 static int nq_m_sync(unsigned int num_gh, struct gfs2_holder *ghs,
1258                      struct gfs2_holder **p)
1259 {
1260         unsigned int x;
1261         int error = 0;
1262
1263         for (x = 0; x < num_gh; x++)
1264                 p[x] = &ghs[x];
1265
1266         sort(p, num_gh, sizeof(struct gfs2_holder *), glock_compare, NULL);
1267
1268         for (x = 0; x < num_gh; x++) {
1269                 p[x]->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1270
1271                 error = gfs2_glock_nq(p[x]);
1272                 if (error) {
1273                         while (x--)
1274                                 gfs2_glock_dq(p[x]);
1275                         break;
1276                 }
1277         }
1278
1279         return error;
1280 }
1281
1282 /**
1283  * gfs2_glock_nq_m - acquire multiple glocks
1284  * @num_gh: the number of structures
1285  * @ghs: an array of struct gfs2_holder structures
1286  *
1287  *
1288  * Returns: 0 on success (all glocks acquired),
1289  *          errno on failure (no glocks acquired)
1290  */
1291
1292 int gfs2_glock_nq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1293 {
1294         struct gfs2_holder *tmp[4];
1295         struct gfs2_holder **pph = tmp;
1296         int error = 0;
1297
1298         switch(num_gh) {
1299         case 0:
1300                 return 0;
1301         case 1:
1302                 ghs->gh_flags &= ~(LM_FLAG_TRY | GL_ASYNC);
1303                 return gfs2_glock_nq(ghs);
1304         default:
1305                 if (num_gh <= 4)
1306                         break;
1307                 pph = kmalloc_array(num_gh, sizeof(struct gfs2_holder *),
1308                                     GFP_NOFS);
1309                 if (!pph)
1310                         return -ENOMEM;
1311         }
1312
1313         error = nq_m_sync(num_gh, ghs, pph);
1314
1315         if (pph != tmp)
1316                 kfree(pph);
1317
1318         return error;
1319 }
1320
1321 /**
1322  * gfs2_glock_dq_m - release multiple glocks
1323  * @num_gh: the number of structures
1324  * @ghs: an array of struct gfs2_holder structures
1325  *
1326  */
1327
1328 void gfs2_glock_dq_m(unsigned int num_gh, struct gfs2_holder *ghs)
1329 {
1330         while (num_gh--)
1331                 gfs2_glock_dq(&ghs[num_gh]);
1332 }
1333
1334 void gfs2_glock_cb(struct gfs2_glock *gl, unsigned int state)
1335 {
1336         unsigned long delay = 0;
1337         unsigned long holdtime;
1338         unsigned long now = jiffies;
1339
1340         gfs2_glock_hold(gl);
1341         holdtime = gl->gl_tchange + gl->gl_hold_time;
1342         if (test_bit(GLF_QUEUED, &gl->gl_flags) &&
1343             gl->gl_name.ln_type == LM_TYPE_INODE) {
1344                 if (time_before(now, holdtime))
1345                         delay = holdtime - now;
1346                 if (test_bit(GLF_REPLY_PENDING, &gl->gl_flags))
1347                         delay = gl->gl_hold_time;
1348         }
1349
1350         spin_lock(&gl->gl_lockref.lock);
1351         handle_callback(gl, state, delay, true);
1352         __gfs2_glock_queue_work(gl, delay);
1353         spin_unlock(&gl->gl_lockref.lock);
1354 }
1355
1356 /**
1357  * gfs2_should_freeze - Figure out if glock should be frozen
1358  * @gl: The glock in question
1359  *
1360  * Glocks are not frozen if (a) the result of the dlm operation is
1361  * an error, (b) the locking operation was an unlock operation or
1362  * (c) if there is a "noexp" flagged request anywhere in the queue
1363  *
1364  * Returns: 1 if freezing should occur, 0 otherwise
1365  */
1366
1367 static int gfs2_should_freeze(const struct gfs2_glock *gl)
1368 {
1369         const struct gfs2_holder *gh;
1370
1371         if (gl->gl_reply & ~LM_OUT_ST_MASK)
1372                 return 0;
1373         if (gl->gl_target == LM_ST_UNLOCKED)
1374                 return 0;
1375
1376         list_for_each_entry(gh, &gl->gl_holders, gh_list) {
1377                 if (test_bit(HIF_HOLDER, &gh->gh_iflags))
1378                         continue;
1379                 if (LM_FLAG_NOEXP & gh->gh_flags)
1380                         return 0;
1381         }
1382
1383         return 1;
1384 }
1385
1386 /**
1387  * gfs2_glock_complete - Callback used by locking
1388  * @gl: Pointer to the glock
1389  * @ret: The return value from the dlm
1390  *
1391  * The gl_reply field is under the gl_lockref.lock lock so that it is ok
1392  * to use a bitfield shared with other glock state fields.
1393  */
1394
1395 void gfs2_glock_complete(struct gfs2_glock *gl, int ret)
1396 {
1397         struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
1398
1399         spin_lock(&gl->gl_lockref.lock);
1400         gl->gl_reply = ret;
1401
1402         if (unlikely(test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags))) {
1403                 if (gfs2_should_freeze(gl)) {
1404                         set_bit(GLF_FROZEN, &gl->gl_flags);
1405                         spin_unlock(&gl->gl_lockref.lock);
1406                         return;
1407                 }
1408         }
1409
1410         gl->gl_lockref.count++;
1411         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1412         __gfs2_glock_queue_work(gl, 0);
1413         spin_unlock(&gl->gl_lockref.lock);
1414 }
1415
1416 static int glock_cmp(void *priv, struct list_head *a, struct list_head *b)
1417 {
1418         struct gfs2_glock *gla, *glb;
1419
1420         gla = list_entry(a, struct gfs2_glock, gl_lru);
1421         glb = list_entry(b, struct gfs2_glock, gl_lru);
1422
1423         if (gla->gl_name.ln_number > glb->gl_name.ln_number)
1424                 return 1;
1425         if (gla->gl_name.ln_number < glb->gl_name.ln_number)
1426                 return -1;
1427
1428         return 0;
1429 }
1430
1431 /**
1432  * gfs2_dispose_glock_lru - Demote a list of glocks
1433  * @list: The list to dispose of
1434  *
1435  * Disposing of glocks may involve disk accesses, so that here we sort
1436  * the glocks by number (i.e. disk location of the inodes) so that if
1437  * there are any such accesses, they'll be sent in order (mostly).
1438  *
1439  * Must be called under the lru_lock, but may drop and retake this
1440  * lock. While the lru_lock is dropped, entries may vanish from the
1441  * list, but no new entries will appear on the list (since it is
1442  * private)
1443  */
1444
1445 static void gfs2_dispose_glock_lru(struct list_head *list)
1446 __releases(&lru_lock)
1447 __acquires(&lru_lock)
1448 {
1449         struct gfs2_glock *gl;
1450
1451         list_sort(NULL, list, glock_cmp);
1452
1453         while(!list_empty(list)) {
1454                 gl = list_entry(list->next, struct gfs2_glock, gl_lru);
1455                 list_del_init(&gl->gl_lru);
1456                 if (!spin_trylock(&gl->gl_lockref.lock)) {
1457 add_back_to_lru:
1458                         list_add(&gl->gl_lru, &lru_list);
1459                         atomic_inc(&lru_count);
1460                         continue;
1461                 }
1462                 if (test_and_set_bit(GLF_LOCK, &gl->gl_flags)) {
1463                         spin_unlock(&gl->gl_lockref.lock);
1464                         goto add_back_to_lru;
1465                 }
1466                 clear_bit(GLF_LRU, &gl->gl_flags);
1467                 gl->gl_lockref.count++;
1468                 if (demote_ok(gl))
1469                         handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1470                 WARN_ON(!test_and_clear_bit(GLF_LOCK, &gl->gl_flags));
1471                 __gfs2_glock_queue_work(gl, 0);
1472                 spin_unlock(&gl->gl_lockref.lock);
1473                 cond_resched_lock(&lru_lock);
1474         }
1475 }
1476
1477 /**
1478  * gfs2_scan_glock_lru - Scan the LRU looking for locks to demote
1479  * @nr: The number of entries to scan
1480  *
1481  * This function selects the entries on the LRU which are able to
1482  * be demoted, and then kicks off the process by calling
1483  * gfs2_dispose_glock_lru() above.
1484  */
1485
1486 static long gfs2_scan_glock_lru(int nr)
1487 {
1488         struct gfs2_glock *gl;
1489         LIST_HEAD(skipped);
1490         LIST_HEAD(dispose);
1491         long freed = 0;
1492
1493         spin_lock(&lru_lock);
1494         while ((nr-- >= 0) && !list_empty(&lru_list)) {
1495                 gl = list_entry(lru_list.next, struct gfs2_glock, gl_lru);
1496
1497                 /* Test for being demotable */
1498                 if (!test_bit(GLF_LOCK, &gl->gl_flags)) {
1499                         list_move(&gl->gl_lru, &dispose);
1500                         atomic_dec(&lru_count);
1501                         freed++;
1502                         continue;
1503                 }
1504
1505                 list_move(&gl->gl_lru, &skipped);
1506         }
1507         list_splice(&skipped, &lru_list);
1508         if (!list_empty(&dispose))
1509                 gfs2_dispose_glock_lru(&dispose);
1510         spin_unlock(&lru_lock);
1511
1512         return freed;
1513 }
1514
1515 static unsigned long gfs2_glock_shrink_scan(struct shrinker *shrink,
1516                                             struct shrink_control *sc)
1517 {
1518         if (!(sc->gfp_mask & __GFP_FS))
1519                 return SHRINK_STOP;
1520         return gfs2_scan_glock_lru(sc->nr_to_scan);
1521 }
1522
1523 static unsigned long gfs2_glock_shrink_count(struct shrinker *shrink,
1524                                              struct shrink_control *sc)
1525 {
1526         return vfs_pressure_ratio(atomic_read(&lru_count));
1527 }
1528
1529 static struct shrinker glock_shrinker = {
1530         .seeks = DEFAULT_SEEKS,
1531         .count_objects = gfs2_glock_shrink_count,
1532         .scan_objects = gfs2_glock_shrink_scan,
1533 };
1534
1535 /**
1536  * examine_bucket - Call a function for glock in a hash bucket
1537  * @examiner: the function
1538  * @sdp: the filesystem
1539  * @bucket: the bucket
1540  *
1541  * Note that the function can be called multiple times on the same
1542  * object.  So the user must ensure that the function can cope with
1543  * that.
1544  */
1545
1546 static void glock_hash_walk(glock_examiner examiner, const struct gfs2_sbd *sdp)
1547 {
1548         struct gfs2_glock *gl;
1549         struct rhashtable_iter iter;
1550
1551         rhashtable_walk_enter(&gl_hash_table, &iter);
1552
1553         do {
1554                 rhashtable_walk_start(&iter);
1555
1556                 while ((gl = rhashtable_walk_next(&iter)) && !IS_ERR(gl))
1557                         if (gl->gl_name.ln_sbd == sdp &&
1558                             lockref_get_not_dead(&gl->gl_lockref))
1559                                 examiner(gl);
1560
1561                 rhashtable_walk_stop(&iter);
1562         } while (cond_resched(), gl == ERR_PTR(-EAGAIN));
1563
1564         rhashtable_walk_exit(&iter);
1565 }
1566
1567 /**
1568  * thaw_glock - thaw out a glock which has an unprocessed reply waiting
1569  * @gl: The glock to thaw
1570  *
1571  */
1572
1573 static void thaw_glock(struct gfs2_glock *gl)
1574 {
1575         if (!test_and_clear_bit(GLF_FROZEN, &gl->gl_flags)) {
1576                 gfs2_glock_put(gl);
1577                 return;
1578         }
1579         set_bit(GLF_REPLY_PENDING, &gl->gl_flags);
1580         gfs2_glock_queue_work(gl, 0);
1581 }
1582
1583 /**
1584  * clear_glock - look at a glock and see if we can free it from glock cache
1585  * @gl: the glock to look at
1586  *
1587  */
1588
1589 static void clear_glock(struct gfs2_glock *gl)
1590 {
1591         gfs2_glock_remove_from_lru(gl);
1592
1593         spin_lock(&gl->gl_lockref.lock);
1594         if (gl->gl_state != LM_ST_UNLOCKED)
1595                 handle_callback(gl, LM_ST_UNLOCKED, 0, false);
1596         __gfs2_glock_queue_work(gl, 0);
1597         spin_unlock(&gl->gl_lockref.lock);
1598 }
1599
1600 /**
1601  * gfs2_glock_thaw - Thaw any frozen glocks
1602  * @sdp: The super block
1603  *
1604  */
1605
1606 void gfs2_glock_thaw(struct gfs2_sbd *sdp)
1607 {
1608         glock_hash_walk(thaw_glock, sdp);
1609 }
1610
1611 static void dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1612 {
1613         spin_lock(&gl->gl_lockref.lock);
1614         gfs2_dump_glock(seq, gl);
1615         spin_unlock(&gl->gl_lockref.lock);
1616 }
1617
1618 static void dump_glock_func(struct gfs2_glock *gl)
1619 {
1620         dump_glock(NULL, gl);
1621 }
1622
1623 /**
1624  * gfs2_gl_hash_clear - Empty out the glock hash table
1625  * @sdp: the filesystem
1626  * @wait: wait until it's all gone
1627  *
1628  * Called when unmounting the filesystem.
1629  */
1630
1631 void gfs2_gl_hash_clear(struct gfs2_sbd *sdp)
1632 {
1633         set_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags);
1634         flush_workqueue(glock_workqueue);
1635         glock_hash_walk(clear_glock, sdp);
1636         flush_workqueue(glock_workqueue);
1637         wait_event_timeout(sdp->sd_glock_wait,
1638                            atomic_read(&sdp->sd_glock_disposal) == 0,
1639                            HZ * 600);
1640         glock_hash_walk(dump_glock_func, sdp);
1641 }
1642
1643 void gfs2_glock_finish_truncate(struct gfs2_inode *ip)
1644 {
1645         struct gfs2_glock *gl = ip->i_gl;
1646         int ret;
1647
1648         ret = gfs2_truncatei_resume(ip);
1649         gfs2_assert_withdraw(gl->gl_name.ln_sbd, ret == 0);
1650
1651         spin_lock(&gl->gl_lockref.lock);
1652         clear_bit(GLF_LOCK, &gl->gl_flags);
1653         run_queue(gl, 1);
1654         spin_unlock(&gl->gl_lockref.lock);
1655 }
1656
1657 static const char *state2str(unsigned state)
1658 {
1659         switch(state) {
1660         case LM_ST_UNLOCKED:
1661                 return "UN";
1662         case LM_ST_SHARED:
1663                 return "SH";
1664         case LM_ST_DEFERRED:
1665                 return "DF";
1666         case LM_ST_EXCLUSIVE:
1667                 return "EX";
1668         }
1669         return "??";
1670 }
1671
1672 static const char *hflags2str(char *buf, u16 flags, unsigned long iflags)
1673 {
1674         char *p = buf;
1675         if (flags & LM_FLAG_TRY)
1676                 *p++ = 't';
1677         if (flags & LM_FLAG_TRY_1CB)
1678                 *p++ = 'T';
1679         if (flags & LM_FLAG_NOEXP)
1680                 *p++ = 'e';
1681         if (flags & LM_FLAG_ANY)
1682                 *p++ = 'A';
1683         if (flags & LM_FLAG_PRIORITY)
1684                 *p++ = 'p';
1685         if (flags & GL_ASYNC)
1686                 *p++ = 'a';
1687         if (flags & GL_EXACT)
1688                 *p++ = 'E';
1689         if (flags & GL_NOCACHE)
1690                 *p++ = 'c';
1691         if (test_bit(HIF_HOLDER, &iflags))
1692                 *p++ = 'H';
1693         if (test_bit(HIF_WAIT, &iflags))
1694                 *p++ = 'W';
1695         if (test_bit(HIF_FIRST, &iflags))
1696                 *p++ = 'F';
1697         *p = 0;
1698         return buf;
1699 }
1700
1701 /**
1702  * dump_holder - print information about a glock holder
1703  * @seq: the seq_file struct
1704  * @gh: the glock holder
1705  *
1706  */
1707
1708 static void dump_holder(struct seq_file *seq, const struct gfs2_holder *gh)
1709 {
1710         struct task_struct *gh_owner = NULL;
1711         char flags_buf[32];
1712
1713         rcu_read_lock();
1714         if (gh->gh_owner_pid)
1715                 gh_owner = pid_task(gh->gh_owner_pid, PIDTYPE_PID);
1716         gfs2_print_dbg(seq, " H: s:%s f:%s e:%d p:%ld [%s] %pS\n",
1717                        state2str(gh->gh_state),
1718                        hflags2str(flags_buf, gh->gh_flags, gh->gh_iflags),
1719                        gh->gh_error,
1720                        gh->gh_owner_pid ? (long)pid_nr(gh->gh_owner_pid) : -1,
1721                        gh_owner ? gh_owner->comm : "(ended)",
1722                        (void *)gh->gh_ip);
1723         rcu_read_unlock();
1724 }
1725
1726 static const char *gflags2str(char *buf, const struct gfs2_glock *gl)
1727 {
1728         const unsigned long *gflags = &gl->gl_flags;
1729         char *p = buf;
1730
1731         if (test_bit(GLF_LOCK, gflags))
1732                 *p++ = 'l';
1733         if (test_bit(GLF_DEMOTE, gflags))
1734                 *p++ = 'D';
1735         if (test_bit(GLF_PENDING_DEMOTE, gflags))
1736                 *p++ = 'd';
1737         if (test_bit(GLF_DEMOTE_IN_PROGRESS, gflags))
1738                 *p++ = 'p';
1739         if (test_bit(GLF_DIRTY, gflags))
1740                 *p++ = 'y';
1741         if (test_bit(GLF_LFLUSH, gflags))
1742                 *p++ = 'f';
1743         if (test_bit(GLF_INVALIDATE_IN_PROGRESS, gflags))
1744                 *p++ = 'i';
1745         if (test_bit(GLF_REPLY_PENDING, gflags))
1746                 *p++ = 'r';
1747         if (test_bit(GLF_INITIAL, gflags))
1748                 *p++ = 'I';
1749         if (test_bit(GLF_FROZEN, gflags))
1750                 *p++ = 'F';
1751         if (test_bit(GLF_QUEUED, gflags))
1752                 *p++ = 'q';
1753         if (test_bit(GLF_LRU, gflags))
1754                 *p++ = 'L';
1755         if (gl->gl_object)
1756                 *p++ = 'o';
1757         if (test_bit(GLF_BLOCKING, gflags))
1758                 *p++ = 'b';
1759         *p = 0;
1760         return buf;
1761 }
1762
1763 /**
1764  * gfs2_dump_glock - print information about a glock
1765  * @seq: The seq_file struct
1766  * @gl: the glock
1767  *
1768  * The file format is as follows:
1769  * One line per object, capital letters are used to indicate objects
1770  * G = glock, I = Inode, R = rgrp, H = holder. Glocks are not indented,
1771  * other objects are indented by a single space and follow the glock to
1772  * which they are related. Fields are indicated by lower case letters
1773  * followed by a colon and the field value, except for strings which are in
1774  * [] so that its possible to see if they are composed of spaces for
1775  * example. The field's are n = number (id of the object), f = flags,
1776  * t = type, s = state, r = refcount, e = error, p = pid.
1777  *
1778  */
1779
1780 void gfs2_dump_glock(struct seq_file *seq, struct gfs2_glock *gl)
1781 {
1782         const struct gfs2_glock_operations *glops = gl->gl_ops;
1783         unsigned long long dtime;
1784         const struct gfs2_holder *gh;
1785         char gflags_buf[32];
1786
1787         dtime = jiffies - gl->gl_demote_time;
1788         dtime *= 1000000/HZ; /* demote time in uSec */
1789         if (!test_bit(GLF_DEMOTE, &gl->gl_flags))
1790                 dtime = 0;
1791         gfs2_print_dbg(seq, "G:  s:%s n:%u/%llx f:%s t:%s d:%s/%llu a:%d v:%d r:%d m:%ld\n",
1792                   state2str(gl->gl_state),
1793                   gl->gl_name.ln_type,
1794                   (unsigned long long)gl->gl_name.ln_number,
1795                   gflags2str(gflags_buf, gl),
1796                   state2str(gl->gl_target),
1797                   state2str(gl->gl_demote_state), dtime,
1798                   atomic_read(&gl->gl_ail_count),
1799                   atomic_read(&gl->gl_revokes),
1800                   (int)gl->gl_lockref.count, gl->gl_hold_time);
1801
1802         list_for_each_entry(gh, &gl->gl_holders, gh_list)
1803                 dump_holder(seq, gh);
1804
1805         if (gl->gl_state != LM_ST_UNLOCKED && glops->go_dump)
1806                 glops->go_dump(seq, gl);
1807 }
1808
1809 static int gfs2_glstats_seq_show(struct seq_file *seq, void *iter_ptr)
1810 {
1811         struct gfs2_glock *gl = iter_ptr;
1812
1813         seq_printf(seq, "G: n:%u/%llx rtt:%llu/%llu rttb:%llu/%llu irt:%llu/%llu dcnt: %llu qcnt: %llu\n",
1814                    gl->gl_name.ln_type,
1815                    (unsigned long long)gl->gl_name.ln_number,
1816                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTT],
1817                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVAR],
1818                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTB],
1819                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SRTTVARB],
1820                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRT],
1821                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_SIRTVAR],
1822                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_DCOUNT],
1823                    (unsigned long long)gl->gl_stats.stats[GFS2_LKS_QCOUNT]);
1824         return 0;
1825 }
1826
1827 static const char *gfs2_gltype[] = {
1828         "type",
1829         "reserved",
1830         "nondisk",
1831         "inode",
1832         "rgrp",
1833         "meta",
1834         "iopen",
1835         "flock",
1836         "plock",
1837         "quota",
1838         "journal",
1839 };
1840
1841 static const char *gfs2_stype[] = {
1842         [GFS2_LKS_SRTT]         = "srtt",
1843         [GFS2_LKS_SRTTVAR]      = "srttvar",
1844         [GFS2_LKS_SRTTB]        = "srttb",
1845         [GFS2_LKS_SRTTVARB]     = "srttvarb",
1846         [GFS2_LKS_SIRT]         = "sirt",
1847         [GFS2_LKS_SIRTVAR]      = "sirtvar",
1848         [GFS2_LKS_DCOUNT]       = "dlm",
1849         [GFS2_LKS_QCOUNT]       = "queue",
1850 };
1851
1852 #define GFS2_NR_SBSTATS (ARRAY_SIZE(gfs2_gltype) * ARRAY_SIZE(gfs2_stype))
1853
1854 static int gfs2_sbstats_seq_show(struct seq_file *seq, void *iter_ptr)
1855 {
1856         struct gfs2_sbd *sdp = seq->private;
1857         loff_t pos = *(loff_t *)iter_ptr;
1858         unsigned index = pos >> 3;
1859         unsigned subindex = pos & 0x07;
1860         int i;
1861
1862         if (index == 0 && subindex != 0)
1863                 return 0;
1864
1865         seq_printf(seq, "%-10s %8s:", gfs2_gltype[index],
1866                    (index == 0) ? "cpu": gfs2_stype[subindex]);
1867
1868         for_each_possible_cpu(i) {
1869                 const struct gfs2_pcpu_lkstats *lkstats = per_cpu_ptr(sdp->sd_lkstats, i);
1870
1871                 if (index == 0)
1872                         seq_printf(seq, " %15u", i);
1873                 else
1874                         seq_printf(seq, " %15llu", (unsigned long long)lkstats->
1875                                    lkstats[index - 1].stats[subindex]);
1876         }
1877         seq_putc(seq, '\n');
1878         return 0;
1879 }
1880
1881 int __init gfs2_glock_init(void)
1882 {
1883         int i, ret;
1884
1885         ret = rhashtable_init(&gl_hash_table, &ht_parms);
1886         if (ret < 0)
1887                 return ret;
1888
1889         glock_workqueue = alloc_workqueue("glock_workqueue", WQ_MEM_RECLAIM |
1890                                           WQ_HIGHPRI | WQ_FREEZABLE, 0);
1891         if (!glock_workqueue) {
1892                 rhashtable_destroy(&gl_hash_table);
1893                 return -ENOMEM;
1894         }
1895         gfs2_delete_workqueue = alloc_workqueue("delete_workqueue",
1896                                                 WQ_MEM_RECLAIM | WQ_FREEZABLE,
1897                                                 0);
1898         if (!gfs2_delete_workqueue) {
1899                 destroy_workqueue(glock_workqueue);
1900                 rhashtable_destroy(&gl_hash_table);
1901                 return -ENOMEM;
1902         }
1903
1904         ret = register_shrinker(&glock_shrinker);
1905         if (ret) {
1906                 destroy_workqueue(gfs2_delete_workqueue);
1907                 destroy_workqueue(glock_workqueue);
1908                 rhashtable_destroy(&gl_hash_table);
1909                 return ret;
1910         }
1911
1912         for (i = 0; i < GLOCK_WAIT_TABLE_SIZE; i++)
1913                 init_waitqueue_head(glock_wait_table + i);
1914
1915         return 0;
1916 }
1917
1918 void gfs2_glock_exit(void)
1919 {
1920         unregister_shrinker(&glock_shrinker);
1921         rhashtable_destroy(&gl_hash_table);
1922         destroy_workqueue(glock_workqueue);
1923         destroy_workqueue(gfs2_delete_workqueue);
1924 }
1925
1926 static void gfs2_glock_iter_next(struct gfs2_glock_iter *gi, loff_t n)
1927 {
1928         struct gfs2_glock *gl = gi->gl;
1929
1930         if (gl) {
1931                 if (n == 0)
1932                         return;
1933                 if (!lockref_put_not_zero(&gl->gl_lockref))
1934                         gfs2_glock_queue_put(gl);
1935         }
1936         for (;;) {
1937                 gl = rhashtable_walk_next(&gi->hti);
1938                 if (IS_ERR_OR_NULL(gl)) {
1939                         if (gl == ERR_PTR(-EAGAIN)) {
1940                                 n = 1;
1941                                 continue;
1942                         }
1943                         gl = NULL;
1944                         break;
1945                 }
1946                 if (gl->gl_name.ln_sbd != gi->sdp)
1947                         continue;
1948                 if (n <= 1) {
1949                         if (!lockref_get_not_dead(&gl->gl_lockref))
1950                                 continue;
1951                         break;
1952                 } else {
1953                         if (__lockref_is_dead(&gl->gl_lockref))
1954                                 continue;
1955                         n--;
1956                 }
1957         }
1958         gi->gl = gl;
1959 }
1960
1961 static void *gfs2_glock_seq_start(struct seq_file *seq, loff_t *pos)
1962         __acquires(RCU)
1963 {
1964         struct gfs2_glock_iter *gi = seq->private;
1965         loff_t n;
1966
1967         /*
1968          * We can either stay where we are, skip to the next hash table
1969          * entry, or start from the beginning.
1970          */
1971         if (*pos < gi->last_pos) {
1972                 rhashtable_walk_exit(&gi->hti);
1973                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
1974                 n = *pos + 1;
1975         } else {
1976                 n = *pos - gi->last_pos;
1977         }
1978
1979         rhashtable_walk_start(&gi->hti);
1980
1981         gfs2_glock_iter_next(gi, n);
1982         gi->last_pos = *pos;
1983         return gi->gl;
1984 }
1985
1986 static void *gfs2_glock_seq_next(struct seq_file *seq, void *iter_ptr,
1987                                  loff_t *pos)
1988 {
1989         struct gfs2_glock_iter *gi = seq->private;
1990
1991         (*pos)++;
1992         gi->last_pos = *pos;
1993         gfs2_glock_iter_next(gi, 1);
1994         return gi->gl;
1995 }
1996
1997 static void gfs2_glock_seq_stop(struct seq_file *seq, void *iter_ptr)
1998         __releases(RCU)
1999 {
2000         struct gfs2_glock_iter *gi = seq->private;
2001
2002         rhashtable_walk_stop(&gi->hti);
2003 }
2004
2005 static int gfs2_glock_seq_show(struct seq_file *seq, void *iter_ptr)
2006 {
2007         dump_glock(seq, iter_ptr);
2008         return 0;
2009 }
2010
2011 static void *gfs2_sbstats_seq_start(struct seq_file *seq, loff_t *pos)
2012 {
2013         preempt_disable();
2014         if (*pos >= GFS2_NR_SBSTATS)
2015                 return NULL;
2016         return pos;
2017 }
2018
2019 static void *gfs2_sbstats_seq_next(struct seq_file *seq, void *iter_ptr,
2020                                    loff_t *pos)
2021 {
2022         (*pos)++;
2023         if (*pos >= GFS2_NR_SBSTATS)
2024                 return NULL;
2025         return pos;
2026 }
2027
2028 static void gfs2_sbstats_seq_stop(struct seq_file *seq, void *iter_ptr)
2029 {
2030         preempt_enable();
2031 }
2032
2033 static const struct seq_operations gfs2_glock_seq_ops = {
2034         .start = gfs2_glock_seq_start,
2035         .next  = gfs2_glock_seq_next,
2036         .stop  = gfs2_glock_seq_stop,
2037         .show  = gfs2_glock_seq_show,
2038 };
2039
2040 static const struct seq_operations gfs2_glstats_seq_ops = {
2041         .start = gfs2_glock_seq_start,
2042         .next  = gfs2_glock_seq_next,
2043         .stop  = gfs2_glock_seq_stop,
2044         .show  = gfs2_glstats_seq_show,
2045 };
2046
2047 static const struct seq_operations gfs2_sbstats_seq_ops = {
2048         .start = gfs2_sbstats_seq_start,
2049         .next  = gfs2_sbstats_seq_next,
2050         .stop  = gfs2_sbstats_seq_stop,
2051         .show  = gfs2_sbstats_seq_show,
2052 };
2053
2054 #define GFS2_SEQ_GOODSIZE min(PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER, 65536UL)
2055
2056 static int __gfs2_glocks_open(struct inode *inode, struct file *file,
2057                               const struct seq_operations *ops)
2058 {
2059         int ret = seq_open_private(file, ops, sizeof(struct gfs2_glock_iter));
2060         if (ret == 0) {
2061                 struct seq_file *seq = file->private_data;
2062                 struct gfs2_glock_iter *gi = seq->private;
2063
2064                 gi->sdp = inode->i_private;
2065                 seq->buf = kmalloc(GFS2_SEQ_GOODSIZE, GFP_KERNEL | __GFP_NOWARN);
2066                 if (seq->buf)
2067                         seq->size = GFS2_SEQ_GOODSIZE;
2068                 /*
2069                  * Initially, we are "before" the first hash table entry; the
2070                  * first call to rhashtable_walk_next gets us the first entry.
2071                  */
2072                 gi->last_pos = -1;
2073                 gi->gl = NULL;
2074                 rhashtable_walk_enter(&gl_hash_table, &gi->hti);
2075         }
2076         return ret;
2077 }
2078
2079 static int gfs2_glocks_open(struct inode *inode, struct file *file)
2080 {
2081         return __gfs2_glocks_open(inode, file, &gfs2_glock_seq_ops);
2082 }
2083
2084 static int gfs2_glocks_release(struct inode *inode, struct file *file)
2085 {
2086         struct seq_file *seq = file->private_data;
2087         struct gfs2_glock_iter *gi = seq->private;
2088
2089         if (gi->gl)
2090                 gfs2_glock_put(gi->gl);
2091         rhashtable_walk_exit(&gi->hti);
2092         return seq_release_private(inode, file);
2093 }
2094
2095 static int gfs2_glstats_open(struct inode *inode, struct file *file)
2096 {
2097         return __gfs2_glocks_open(inode, file, &gfs2_glstats_seq_ops);
2098 }
2099
2100 static int gfs2_sbstats_open(struct inode *inode, struct file *file)
2101 {
2102         int ret = seq_open(file, &gfs2_sbstats_seq_ops);
2103         if (ret == 0) {
2104                 struct seq_file *seq = file->private_data;
2105                 seq->private = inode->i_private;  /* sdp */
2106         }
2107         return ret;
2108 }
2109
2110 static const struct file_operations gfs2_glocks_fops = {
2111         .owner   = THIS_MODULE,
2112         .open    = gfs2_glocks_open,
2113         .read    = seq_read,
2114         .llseek  = seq_lseek,
2115         .release = gfs2_glocks_release,
2116 };
2117
2118 static const struct file_operations gfs2_glstats_fops = {
2119         .owner   = THIS_MODULE,
2120         .open    = gfs2_glstats_open,
2121         .read    = seq_read,
2122         .llseek  = seq_lseek,
2123         .release = gfs2_glocks_release,
2124 };
2125
2126 static const struct file_operations gfs2_sbstats_fops = {
2127         .owner   = THIS_MODULE,
2128         .open    = gfs2_sbstats_open,
2129         .read    = seq_read,
2130         .llseek  = seq_lseek,
2131         .release = seq_release,
2132 };
2133
2134 int gfs2_create_debugfs_file(struct gfs2_sbd *sdp)
2135 {
2136         struct dentry *dent;
2137
2138         dent = debugfs_create_dir(sdp->sd_table_name, gfs2_root);
2139         if (IS_ERR_OR_NULL(dent))
2140                 goto fail;
2141         sdp->debugfs_dir = dent;
2142
2143         dent = debugfs_create_file("glocks",
2144                                    S_IFREG | S_IRUGO,
2145                                    sdp->debugfs_dir, sdp,
2146                                    &gfs2_glocks_fops);
2147         if (IS_ERR_OR_NULL(dent))
2148                 goto fail;
2149         sdp->debugfs_dentry_glocks = dent;
2150
2151         dent = debugfs_create_file("glstats",
2152                                    S_IFREG | S_IRUGO,
2153                                    sdp->debugfs_dir, sdp,
2154                                    &gfs2_glstats_fops);
2155         if (IS_ERR_OR_NULL(dent))
2156                 goto fail;
2157         sdp->debugfs_dentry_glstats = dent;
2158
2159         dent = debugfs_create_file("sbstats",
2160                                    S_IFREG | S_IRUGO,
2161                                    sdp->debugfs_dir, sdp,
2162                                    &gfs2_sbstats_fops);
2163         if (IS_ERR_OR_NULL(dent))
2164                 goto fail;
2165         sdp->debugfs_dentry_sbstats = dent;
2166
2167         return 0;
2168 fail:
2169         gfs2_delete_debugfs_file(sdp);
2170         return dent ? PTR_ERR(dent) : -ENOMEM;
2171 }
2172
2173 void gfs2_delete_debugfs_file(struct gfs2_sbd *sdp)
2174 {
2175         if (sdp->debugfs_dir) {
2176                 if (sdp->debugfs_dentry_glocks) {
2177                         debugfs_remove(sdp->debugfs_dentry_glocks);
2178                         sdp->debugfs_dentry_glocks = NULL;
2179                 }
2180                 if (sdp->debugfs_dentry_glstats) {
2181                         debugfs_remove(sdp->debugfs_dentry_glstats);
2182                         sdp->debugfs_dentry_glstats = NULL;
2183                 }
2184                 if (sdp->debugfs_dentry_sbstats) {
2185                         debugfs_remove(sdp->debugfs_dentry_sbstats);
2186                         sdp->debugfs_dentry_sbstats = NULL;
2187                 }
2188                 debugfs_remove(sdp->debugfs_dir);
2189                 sdp->debugfs_dir = NULL;
2190         }
2191 }
2192
2193 int gfs2_register_debugfs(void)
2194 {
2195         gfs2_root = debugfs_create_dir("gfs2", NULL);
2196         if (IS_ERR(gfs2_root))
2197                 return PTR_ERR(gfs2_root);
2198         return gfs2_root ? 0 : -ENOMEM;
2199 }
2200
2201 void gfs2_unregister_debugfs(void)
2202 {
2203         debugfs_remove(gfs2_root);
2204         gfs2_root = NULL;
2205 }